Nvidia expanded its safety focus beyond autonomous vehicles to physical AI, launching Halos for Robotics to ensure safe deployments of mobile robots and humanoids. This safety system integrates hardware, such as specialized computing modules, and software to monitor and isolate safety-critical operations. The Halos system includes tools for monitoring hardware blocks, connecting sensor data, and running simulations for testing in virtual environments. Companies like Agility Robotics and Boston Dynamics are incorporating Halos to enable robots to operate safely in diverse, unstructured physical settings. This initiative aims to enable robots to handle multiple tasks by improving safety across various applications.
Humanoid robots have demonstrated impressive athletic abilities like sprinting and boxing in competitions. However, the true test of their readiness for everyday life lies in complex tasks such as doing the laundry. Robots struggle with the unpredictable nature of folding clothes due to varying shapes, materials, and creases. Successfully learning these skills requires robots to adapt to reality rather than just replicating pre-programmed movements. Therefore, the ability to handle messy, real-world chores may be a better indicator of general-purpose robot intelligence than athletic feats.
Embodied coding agents generate robot policies from natural language, but execution feedback is insufficient for aerial manipulation due to partial observability and fragile physical interactions. The paper introduces Active Feedback (AF), a "View and Probe" framework, to address this limitation. AF allows agents to synthesize diagnostic programs to inspect unseen outcomes or test failure hypotheses through targeted physical interactions. Across simulations, AF significantly reduces false completion declarations and increases task success rates compared to terminal feedback. This work demonstrates that executable diagnosis effectively fills gaps in execution feedback for embodied agents in aerial manipulation.
PanTact is a system designed to teach humanoid robots whole-body dexterity by learning from robot-free human demonstrations. It utilizes a wearable, robot-free interface with modular contact patches to capture body contact locations and states as manipulation-centric information. A visuomotor policy processes this data to predict whole-body contact goals despite visual limitations. Contact-aware control formulations then execute these goals by coordinating the robot's body to interact with suitable surfaces. The system has been demonstrated on a Unitree G1 humanoid, enabling actions like pushing and load-sharing.
A benchmark called MeMoREval is introduced to study memory modality for embodied agents by using a controlled experience-request protocol. Agents experience events first and then use memory to respond to requests without new observations. The benchmark involves 53 tasks across four simulated indoor environments, measuring success and inference costs. Evaluations suggest storing frames outperforms storing descriptions for lower write-time costs, while storing frames with scene graphs yields the highest task success. The work also explores extensions to accommodate further memory design investigations.
The paper formalizes zero-shot manipulation from synthetic videos by treating it as a constrained optimization problem. It addresses the issue that synthetic videos lack physical consistency and robot grounding, preventing direct trajectory rollout. The method uses object trajectories from the video as an initial guess for optimization within a digital twin. The formalism shows that video generation must probabilistically seed the search within a semantic basin of attraction for successful manipulation. Experiments demonstrate that removing simplifications leads to more robust zero-shot trajectories and improved task success rates.
Motion planning for robots in cluttered environments is complicated by treating all objects as fixed obstacles. The Minimum Constraint Removal (MCR) problem seeks the fewest movable object removals necessary for a collision-free path, which is NP-hard. This work proposes a graph-based heuristic pipeline combining feasible-set generation and fixed-length searches over reduced MCR graphs. Evaluation across six benchmarks showed the method achieved optimal solutions in some cases and provided significant speedups over exact search. The approach balances computational efficiency with solution quality for motion planning among movable obstacles.
A 3D path-planning algorithm, extending the A* algorithm, is proposed for navigating soft-growing robots in the spinal subarachnoid space. This method aims to estimate safe pathways while establishing anchor points to facilitate efficient robot steering. The approach considers a highly detailed model of the spinal subarachnoid space for evaluation. The research focuses on ensuring the estimated paths satisfy the robot's movement requirements and path efficiency. The work is relevant to developing navigation strategies for minimally invasive soft robotic procedures.
An electrothermally actuated fiber-based robot is presented for high-precision minimally invasive surgery. Joule heating is concentrated in the distal active segment of the fiber to achieve precise distal manipulation. The fiber robot, measuring 1.9 mm in diameter and 1.2 m in length, can position surgical laser fibers with accuracy below 100 $\mu$m. This compact design allows the robot to pass through an endoscope working channel while remaining mechanically uncoupled. The system demonstrates potential to improve clinical outcomes by enhancing precision in MIS procedures.
The research proposes a distributed harmonic-covariance feedback method to achieve hole closure in planar robot swarms. This method utilizes harmonic edge signals to identify holes and converts these into agent forces for distributed evaluation. A fixed-degree polynomial of the regularized Hodge Laplacian is used for distributed assessment via local operator applications. Offline calculations yield rational lower bounds for cumulative hole area, which converge to the actual area, indicating hole closure over time. Numerical demonstrations show that holes close within a short time frame.
Motion retargeting for humanoid robots requires handling complex spatial relationships beyond simple motion transfer. PriMoRetarget introduces a framework that replaces relative objective weighting with an explicit priority hierarchy for motion retargeting. This allows critical elements like foot placement and interaction contacts to be prioritized over less important pose objectives. The method is evaluated across datasets involving single-character motion, object manipulation, and character interactions. Consequently, the resulting motions better preserve task-relevant interactions and provide more reliable reference trajectories.
The research addresses planning for multiple UAV-UGV teams in aerial monitoring missions with strict ordering requirements for monitoring points. The system incorporates UAV energy limitations, modeled by flight endurance, and utilizes UGVs as mobile charging stations and transport. The objective is to minimize mission completion time while respecting energy constraints and the required visit sequence, even with travel-time disturbances. A novel polynomial-time algorithm is developed to jointly determine vehicle routes and timing, allowing for dynamic adjustments to meet ordering constraints. The work provides feasibility guarantees, complexity analysis, and performance evaluations for the proposed planning method.